SAN FRACISCO >> Pacific Gas & Electric Company is using AI and machine learning tools to create more accurate forecasts which rely on real-time information from a network of weather stations as well as historical data to keep people safe during wildfire season, typically May through October (longer along dry shoreline).
In a podcast Thursday from PG&E headquarters, Scott Strenfel, meteorologist who runs the fire science teams at PG&E, explained how its newest digital tools installed in 1,600 weather stations throughout California are providing real-time data to make more accurate weather forecasts so to take appropriate action. Strenfel noted temperatures have been cooler than normal this year, yet wildfires still pop up. In 2024 there were 5,132 wildfires while this year there have been 5,881 wildfires. Last year, 40,824 acres burned while this year 387,002 acres burned. This year it has been the coolest temperatures since 2011. "That's due to the Marine Layer offshore and a stagnant low - pressure area off the Pacific-Northwest. "But we'll have fire events into September even October," he said.
A big heat event followed by a Diablo Wind event can bring more wildfire activity. There have been more wildfires reported by Cal Fire, the Forest Service and BLM compared to last year and compared to the 5-year average. "California has fire activity that's still significant, one of the first mega-fires being the Gifford Fire in Santa Barbara, over 100,000 acres burned. "What data is showing for the next few months is above normal wildfire potential, but Southern California fire danger sticks around because of drought conditions," he said. This past winter was drier than normal across Southern California, and it is a similar pattern as last winter. PG&E aims to minimize outage events as much as possible with advance situational analysis awareness.
It includes a base map which covers 70,000 square miles of California (Santa Barbara to almost Cresent City) and from sea level to Kirkwood Mountain. The topography varies widely and makes monitoring difficult with all the microclimates that makes accurate forecasts sometimes difficult. Since 2017, Automated Surface Observing Systems, installed by Cal Fire at area airports, help to understand where there are Diablo Wind events that are across California. But it has led to an advanced situational awareness and PG&E is still installing stations in strategic locations which had been hard to do because of permitting regulations.
Also helping advance data collection are machine learning (computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyze and draw inferences from patterns in data). "We've developed a machine learning model based on how each individual weather station performs compared to high resolution models PG&E has," he said. "Essentially we are able to improve the weather forecast ... We get a wind forecast from anyone of PG&E's weather stations and starts to show me where wind gusts exceed 40 MPH. And we get a variety of other observations and quickly see which weather stations are going to be the windiest over the next five days. Even with 1,600 stations there's locations between where you are not going to measure- so we have Real Time Emphasis Scale Analysis."
It takes all those observations, then run them through a model and provides meteorologists with a grided view of, say, what the wind speed looks like and see where winds are the strongest that start to move down the San Joaquin Valley. "That's the weather visualization, and we can take that real time layer and run it through our machine learning models, such as the Wildfire Potential Index WFPI, developed using an algorithm that describes the ratio of live to dead fuels and includes variables for wind speed, dry bulb temperature and rainfall," Strenfel said. "And the WFPI is one of the most crucial models in the company. It can predict the probability if we were to have an ignition event in one of these small locations and what's the probability that ignition becomes a larger catastrophic fire from any ignition source."
There are Native American scientists on a team who run data through an AI algorithm and then through a learning machine so technologists can see in good detail about wildfire risk across regional territory. A World Face Map, illustrates a current state of atmosphere over a larger region, displaying various weather parameters; wind speed and direction, humidity, all essential to understand weather patterns so to forecast across different regions. So, one of the map layers that went into the WFPI shows the kinds of fuels around an area, whether it is urban, suburb, grassed, shrub or body of water.
In addition, meteorologists use an outage probability weather model combined with data from the WFPI model, as a decision-making framework to determine if they should execute a Public Safety Power Shutoff event, as well as the locations of where winds might take down trees or infrastructure apparatus. And all of this data is fed to computers at a command or emergency operations center to determine if a PSPS event should be executed.
At this point Strenfel was joined by Andria Borba, communications representative to conduct a brief question and answer session. Borba noted extreme and shifting weather phenomena is real. "The climate is shifting rapidly," Borba said. "But are these tools make individuals- whole towns safer?"
"Absolutely, this situational awareness platform provides us with unbelievable intelligence of what's happening across those 70,00 square miles that we never have had before and will only continue to get better." But he also acknowledged Borbas's point, that climate change is increasing risk too. Strenfel stressed the need for Californians to raise their level of preparedness. Have evacuation plans ready. "Fires can happen any point in time," he said. "Have a to-go bag ready, practice your evacuation plan and communicate it to loved ones."